From: Edward Cree <ec...@solarflare.com>
Date: Wed, 17 May 2017 15:00:04 +0100
> On 16/05/17 23:53, Alexei Starovoitov wrote:
>> following this line of thinking it feels that it should be possible
>> to get rid of 'aux_off' and 'aux_off_align' and simplify the code.
>> I mean we can always do
>> dst_reg->min_align = min(dst_reg->min_align, src_reg->min_align);
>>
>> and don't use 'off' as part of alignment checks at all.
> Problem with this approach, of course, is that (say) NET_IP_ALIGN +
> sizeof(ethhdr) = 16 is muchly aligned, whereas if you turn all
> constants into alignments you think you're only 2-byte aligned.
> I think you have to track exact offsets when you can, and only turn
> into an alignment when you introduce a variable.
> Of course it can still be fooled by e.g. 2 + (x << 2) + 14, which it
> will think is only 2-aligned when really it's 4-aligned, but unless
> you want to start tracking 'bits known to be 1' as well as 'bits
> known to be 0', I think you just accept that alignment tracking
> isn't commutative. The obvious cases (ihl << 2 and so) will work
> when written the obvious way, unless the compiler does something
> perverse.
> OTOH the 'track known 1s as well' might work in a nice generic way
> and cover all bases, I'll have to experiment a bit with that.
Both cases are common in real BPF programs. The offsets really are
necessary. It's funny because initially I tried to implement this
without the auxiliary offset and it simply doesn't work. :-)
We always have to track when you've seen the offset that cancels out
the NET_IP_ALIGN. And as stated it can occur both before and after
variable offsets have been introduced.
You have to catch both:
ptr += variable;
ptr += 14;
and:
ptr += 14;
ptr += variable; /* align = 4 */
And always see at the end that "NET_IP_ALIGN + offsets" will
be properly 4 byte aligned.
